How Shared Embodied Intelligence Redefines Human‑Robot Collaboration

Shared Embodied Intelligence (SEI) Framework

The SEI framework jointly optimises a humanoid robot’s mechanical structure, control hierarchy and a digital human biomechanical model within a single optimisation loop. The human model includes skeleton, joints and muscle dynamics, enabling real‑time estimation of the human’s force state and motion trajectory.

The robot (ergoCub) uses a hierarchical control architecture that links high‑level task planning to low‑level whole‑body control, forming a closed‑loop system that continuously exchanges information with the human model.

A symmetric information‑exchange mechanism lets the estimated human state continuously influence the robot’s motion planning, while the robot’s motions in turn modify the human’s load distribution. The optimisation objective therefore minimises a combined “human‑robot” cost that includes balance, efficiency, safety and comfort, rather than robot balance alone.

Co‑design and Optimisation

Robot hardware parameters, control gains and the human biomechanical model are embedded in the same optimisation problem. The optimisation seeks to reduce mechanical load on the human, explicitly targeting the L5‑S1 lumbar joint, a known injury‑prone location during manual handling.

Experimental Validation

Collaborative Lifting Task In a collaborative lifting experiment, the robot continuously estimated the human’s posture and computed lumbar joint torque. The computed lumbar load was added to the robot’s cost function, causing the robot to adjust its posture, force direction and movement rhythm to lower the human’s mechanical burden. Figure 3 (original article) shows the reduction in lumbar torque compared with a baseline that ignores human load.

External‑Force Disturbance Task External pushes were applied to the robot to simulate collisions, disturbances or sudden human movements. Instead of rigidly following a predefined trajectory, the robot re‑balanced its whole body based on the updated human state, maintaining collaborative stability while tracking the desired path. Figure 4 (original article) illustrates the controller’s tracking performance under disturbance.

Both experiments demonstrate that embedding a human biomechanical model into the robot’s design and control loop allows the robot to act as a true collaborative partner, improving safety and efficiency.

Reference: "Towards shared embodied intelligence in humanoid robots through optimisation, development and testing of the human‑aware ergoCub robot", Nature Machine Intelligence , https://www.nature.com/articles/s42256-026-01272-2

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
机器人不单独学习如何运动，而是与人体一起被建模、一起被优化。